The objective of Task 9 (Safety Warning Countermeasures) is to improve safety warning systems by designing these systems to ad

5.6.3.Discussion

The performance of BN models confirmed that they are a useful and promising technique for identifying driver cognitive distraction. Drivers behaved differently when they were cognitively loaded and BNs were able to detect these differences. Eye movement and driving performance measures, especially those having high mutual information with distraction, were useful in predicting drivers’ cognitive state. The best predictors of distraction were DBNs using 120-second training sequences. These models produced an accuracy of 86.4%, with a sensitivity of 3.90. DBNs produced more sensitive models, suggesting that time-dependent data contains information that indicates distraction.

The presence of a hidden node, particularly for the DBN, diminished model performance. This was surprising in that a hidden node represents a free parameter that should enable the model to fit the data more precisely. One interpretation of this result is that including an intermediate state in the model does not accurately describe the mechanisms underlying cognitive distraction. The model structure without a hidden node is a more accurate representation of the relationship between driver performance and cognitive state. Another possibility is that, because hidden nodes introduce more uncertainty in the model learning process, using equal amounts of training data for the models both with and without hidden nodes may have caused the hidden node models to be less accurate. This second possibility would also explain why accuracy decreased from no hidden node to one hidden node more for the DBNs than for the SBNs; the learning process was more complex for DBNs. To determine which explanation is correct, another study with is needed to ensure that training can produce accurate parameter estimates.

Window size did not affect model performance for either the SBNs or DBNs, which means that evidence summarized across different periods of time did not present the distraction signal differently, and that the definition of time step did not affect model performance of DBNs. These results conflict with those from Liang et al. (Liang et al.), where larger windows improved the ability of SVM to detect cognitive distraction. The conflicting results may reflect important differences between the two methods. As with the hidden node comparison, the models that use longer window sizes were trained with fewer instances because the length of total data was fixed, meaning that large windows produced fewer training instances. The effect of sequential length of training examples conforms to this explanation, with sequential length of training examples improving DBNs model performance. Longer sequences supply more repeated cases from which DBNs can extract the stable relationship between performance and cognitive distraction.

The analysis of mutual information showed that blink frequency had the strongest relationship with cognitive distraction. Another study used blink measures to study driver vigilance (Bergasa et al., 2006), and found that when drivers were alert, they blinked less frequently. We found a similar pattern of behavior; during baseline drives, drivers blinked at 0.31Hz, but when they interacted with the secondary task, they blinked at 0.49 Hz (F(2,18) = 361.4, p<0.0001). Such an increase in involuntary eye movements may disrupt the consolidation of visual information (Strayer et al., 2003b). The time-dependent changes in fixation horizontal distribution, pursuit speed, and driving measures are all more predictive of distraction than when such changes are considered as static measures, because their mutual information in DBN inter-structure (5%, 3%, and 2%, respectively) was higher than that in SBN structure (2%, 1%, and 0.5%, respectively). The mutual information analysis also shows that individual differences render unrealistic a uniform model structure that fits all people.

Interestingly, larger window sizes decreased the number of links and increased the mutual information per link in the trained models. This suggests that larger window sizes make important dependent relationships between performance measures and distraction more prominent, but do not necessarily improve model performance. One possible reason is that training instances decrease as window size increases.

Although this BN approach aims at real-time detection, there remains an unavoidable lag from when the driver’s cognitive state changes until the model recognizes the change. As with the SVM algorithm, the lag results largely from three sources: sensor delay, model computational delay, and the delay caused by summarizing the measures across a window. Sensor delay is the time required for data acquisition and reduction, and model computational delay is the computational time of BN models. The magnitudes of the first two kinds of delay are on the order of seconds. The third source reflects the need to summarize the measures across a window, which is expected to be approximately half of window size. That is, a 5-second window produces about a 2.5-second delay, and 30-second windows produce about 15-second delays. As a consequence, increasing window size also increases model lag. The cost of the lag needs to be balanced against the prediction accuracy needed to support the specific distribution mitigation strategy. For the BNs in this study, the greater number of training instances seemed to balance the less precise state estimates associated with small time windows. Further analysis of larger datasets with more training instances are needed to assess this possibility and the tradeoff between prediction accuracy and prediction lag.

Although using BN to identify driver distraction represents a promising approach, implementation in production vehicles will depend on the development of sensors. For example, economical eye tracking devices that are robust to road surface, lighting conditions, eye color, and eye glasses are only now becoming a reality. The outcomes of this study suggest that as eye tracking and other driver-state sensors become more available, real-time measurement of driver distraction may be feasible. BN models that identify driver distraction provide critical input to distraction mitigation strategies that might include warning a distracted driver to attend to the road, diminishing the threshold of collision warnings, or recording instances of distraction for later review (Donmez et al., 2003a).